Bundle adjustment#

Work in progress

Placeholder content. Being rewritten with figures.

The vendor’s camera models are slightly inaccurate; bundle adjustment refines them by minimizing reprojection errors of features matched between input images.

What the optimizer does#

Given initial cameras and tie points, bundle_adjust jointly varies camera parameters and 3D tie-point positions to drive reprojection error toward zero.

Why this matters for stereo#

Without bundle adjustment, vendor camera misalignment shows up as bias and tilt in the final DEM; with it, residual reprojection error and DEM bias both shrink.

Outputs you can visualize#

Initial- and final-pass per-tie-point residual CSVs (*-initial_residuals_pointmap.csv, *-final_residuals_pointmap.csv) are read by asp_plot.bundle_adjust.PlotBundleAdjustFiles for before/after comparison.

Common knobs#

Interest-point density (--ip-per-image), tie-point penalty (--tri-weight, --tri-robust-threshold), and camera-anchor weight (--camera-weight) are the parameters most often tuned.

Where to read more#